
Editorial illustration.
AI Paper Screening: Check the Decision Before Applying It
INRA.AI Team
AI Research Platform
An AI screening suggestion is a proposed decision about a record. Before using it, check which eligibility criterion it addresses, what text was available and who will make the final decision.
For clinical research teams, the cost of a mistaken exclusion can be a missing study, not just an untidy reading list. This guide concerns systematic-review screening. Finding useful papers for a narrative review is a different task and does not establish that a systematic search or selection process has been completed.
Separate a promising abstract from an eligible study
A title or abstract can suggest relevance without resolving every criterion. A record with no abstract leaves less information to assess. Full-text retrieval can also fail or return an incomplete document. Keep these states visible; “not available to the tool” does not mean “not reported by the study.”
Cochrane chapter 4, section 4.6.4 distinguishes initial screening from final eligibility assessment. It calls for at least two people working independently on the final full-text eligibility decision. A high AI confidence score does not replace that requirement.
Make the reason inspectable
For each proposed exclusion, examine the criterion, the source passage and any missing information. These are practical review prompts, not a validated screening instrument:
- Population and setting: does the report actually describe an ineligible group, or is the abstract simply unclear?
- Intervention and comparison: is the relevant component absent, or described using different terminology?
- Study design: is the classification supported by the methods, rather than inferred from a title?
- Outcome: distinguish an outcome not mentioned in the abstract from one that was not measured. Apply the eligibility rules your team agreed.
Keep a reason tied to the protocol. A broad label such as “not relevant” gives a second reviewer little to inspect. When the report does not settle a criterion, follow the team's planned process for obtaining information and resolving uncertainty.
Understand what INRA's controls do
SLR Copilot has title/abstract queues, reviewer votes, AI suggestions, conflict resolution and a separate full-text phase. It also includes bulk-voting and bulk-apply controls. A reviewer initiating a bulk action is not evidence that each suggestion received an independent human assessment.
Before applying suggestions to a group, inspect the selected records, filters, decision and recorded reason. Evaluate errors with the methods team and document the chosen process. This guide does not establish a safe sample size, an acceptable miss rate or a time saving.
Automatic PDF retrieval is limited by access and availability. Review the retrieval status and use permitted manual or institutional access when needed. An accessible PDF is not evidence that a study is eligible or reliable.
Keep records, reports and studies distinct
Two database records may describe the same report; two different reports may describe the same study. Deduplicating records does not settle the second relationship. Cochrane chapter 5 explains why linked reports matter when collecting study data. Keep identifiers and explanations available as the review progresses.
Report the decisions that actually happened
PRISMA 2020 supports reporting. Reconcile counts, exclusions and any automation description with the underlying records. A diagram does not certify search completeness, selection accuracy or methodological quality.
The screening controls described here belong to SLR Copilot. Do not infer that they are part of INRA's separate NLR generation workflow.
Updated October 19, 2026 to clarify product limits and source checks. Qualified methods review of this update is pending.